The document proposes a new metaheuristic optimization algorithm called the Bat Algorithm (BA) which is inspired by the echolocation behavior of microbats. The BA idealizes how bats use echolocation to locate prey and avoid obstacles. It formulates rules to govern how virtual "bats" change their pulse frequencies, loudness, and locations to search for the optimal solution. Simulations show the BA performs well on benchmark test functions, finding the global minimum solutions. The author believes combining advantages of existing algorithms led to developing this new BA which seems superior and warrants further study.

1. arXiv:1004.4170v1[math.OC]23Apr2010
A New Metaheuristic Bat-Inspired Algorithm
Xin-She Yang
Department of Engineering, University of Cambridge,
Trumpington Street, Cambridge CB2 1PZ, UK
Abstract
Metaheuristic algorithms such as particle swarm optimization, firefly algorithm and
harmony search are now becoming powerful methods for solving many tough optimiza-
tion problems. In this paper, we propose a new metaheuristic method, the Bat Algo-
rithm, based on the echolocation behaviour of bats. We also intend to combine the
advantages of existing algorithms into the new bat algorithm. After a detailed formu-
lation and explanation of its implementation, we will then compare the proposed algo-
rithm with other existing algorithms, including genetic algorithms and particle swarm
optimization. Simulations show that the proposed algorithm seems much superior to
other algorithms, and further studies are also discussed.
Citation detail:
X.-S. Yang, A New Metaheuristic Bat-Inspired Algorithm, in: Nature Inspired Coop-
erative Strategies for Optimization (NISCO 2010) (Eds. J. R. Gonzalez et al.), Studies
in Computational Intelligence, Springer Berlin, 284, Springer, 65-74 (2010).
1 Introduction
Metaheuristic algorithms such as particle swarm optimization and simulated annealing
are now becoming powerful methods for solving many tough optimization problems [3-
7,11]. The vast majority of heuristic and metaheuristic algorithms have been derived
from the behaviour of biological systems and/or physical systems in nature. For exam-
ple, particle swarm optimization was developed based on the swarm behaviour of birds
and fish [7, 8], while simulated annealing was based on the annealing process of metals
[9].
New algorithms are also emerging recently, including harmony search and the firefly
algorithm. The former was inspired by the improvising process of composing a piece
of music [4], while the latter was formulated based on the flashing behaviour of fireflies
[14]. Each of these algorithms has certain advantages and disadvantages. For example,
simulating annealing can almost guarantee to find the optimal solution if the cooling
process is slow enough and the simulation is running long enough; however, the fine
adjustment in parameters does affect the convergence rate of the optimization process.
A natural question is whether it is possible to combine major advantages of these
algorithms and try to develop a potentially better algorithm? This paper is such an
attempt to address this issue.
1

2. In this paper, we intend to propose a new metaheuristic method, namely, the Bat
Algorithm (BA), based on the echolocation behaviour of bats. The capability of echolo-
cation of microbats is fascinating as these bats can find their prey and discriminate
different types of insects even in complete darkness. We will first formulate the bat
algorithm by idealizing the echolocation behaviour of bats. We then describe how it
works and make comparison with other existing algorithms. Finally, we will discuss
some implications for further studies.
2 Echolocation of bats
2.1 Behaviour of microbats
Bats are fascinating animals. They are the only mammals with wings and they also have
advanced capability of echolocation. It is estimated that there are about 996 different
species which account for up to 20% of all mammal species [1, 2]. Their size ranges from
the tiny bumblebee bat (of about 1.5 to 2g) to the giant bats with wingspan of about 2
m and weight up to about 1 kg. Microbats typically have forearm length of about 2.2 to
11cm. Most bats uses echolocation to a certain degree; among all the species, microbats
are a famous example as microbats use echolocation extensively while megabats do not
[12, 13].
Most microbats are insectivores. Microbats use a type of sonar, called, echolocation,
to detect prey, avoid obstacles, and locate their roosting crevices in the dark. These bats
emit a very loud sound pulse and listen for the echo that bounces back from the sur-
rounding objects. Their pulses vary in properties and can be correlated with their hunt-
ing strategies, depending on the species. Most bats use short, frequency-modulated sig-
nals to sweep through about an octave, while others more often use constant-frequency
signals for echolocation. Their signal bandwidth varies depends on the species, and
often increased by using more harmonics.
2.2 Acoustics of Echolocation
Though each pulse only lasts a few thousandths of a second (up to about 8 to 10 ms),
however, it has a constant frequency which is usually in the region of 25kHz to 150
kHz. The typical range of frequencies for most bat species are in the region between
25kHz and 100kHz, though some species can emit higher frequencies up to 150 kHz.
Each ultrasonic burst may last typically 5 to 20 ms, and microbats emit about 10 to 20
such sound bursts every second. When hunting for prey, the rate of pulse emission can
be sped up to about 200 pulses per second when they fly near their prey. Such short
sound bursts imply the fantastic ability of the signal processing power of bats. In fact,
studies shows the integration time of the bat ear is typically about 300 to 400 µs.
As the speed of sound in air is typically v = 340 m/s, the wavelength λ of the
ultrasonic sound bursts with a constant frequency f is given by
λ =
v
f
, (1)
which is in the range of 2mm to 14mm for the typical frequency range from 25kHz to
150 kHz. Such wavelengths are in the same order of their prey sizes.
2

3. Amazingly, the emitted pulse could be as loud as 110 dB, and, fortunately, they
are in the ultrasonic region. The loudness also varies from the loudest when searching
for prey and to a quieter base when homing towards the prey. The travelling range of
such short pulses are typically a few metres, depending on the actual frequencies [12].
Microbats can manage to avoid obstacles as small as thin human hairs.
Studies show that microbats use the time delay from the emission and detection of
the echo, the time difference between their two ears, and the loudness variations of the
echoes to build up three dimensional scenario of the surrounding. They can detect the
distance and orientation of the target, the type of prey, and even the moving speed of
the prey such as small insects. Indeed, studies suggested that bats seem to be able to
discriminate targets by the variations of the Doppler effect induced by the wing-flutter
rates of the target insects [1].
Obviously, some bats have good eyesight, and most bats also have very sensitive
smell sense. In reality, they will use all the senses as a combination to maximize the
efficient detection of prey and smooth navigation. However, here we are only interested
in the echolocation and the associated behaviour.
Such echolocation behaviour of microbats can be formulated in such a way that it
can be associated with the objective function to be optimized, and this make it possible
to formulate new optimization algorithms. In the rest of this paper, we will first outline
the basic formulation of the Bat Algorithm (BA) and then discuss the implementation
and comparison in detail.
3 Bat Algorithm
If we idealize some of the echolocation characteristics of microbats, we can develop var-
ious bat-inspired algorithms or bat algorithms. For simplicity, we now use the following
approximate or idealized rules:
1. All bats use echolocation to sense distance, and they also ‘know’ the difference
between food/prey and background barriers in some magical way;
2. Bats fly randomly with velocity vi at position xi with a fixed frequency fmin,
varying wavelength λ and loudness A0 to search for prey. They can automatically
adjust the wavelength (or frequency) of their emitted pulses and adjust the rate
of pulse emission r ∈ [0, 1], depending on the proximity of their target;
3. Although the loudness can vary in many ways, we assume that the loudness varies
from a large (positive) A0 to a minimum constant value Amin.
Another obvious simplification is that no ray tracing is used in estimating the time
delay and three dimensional topography. Though this might be a good feature for
the application in computational geometry, however, we will not use this as it is more
computationally extensive in multidimensional cases.
In addition to these simplified assumptions, we also use the following approxima-
tions, for simplicity. In general the frequency f in a range [fmin, fmax] corresponds to
a range of wavelengths [λmin, λmax]. For example a frequency range of [20kHz, 500kHz]
corresponds to a range of wavelengths from 0.7mm to 17mm.
For a given problem, we can also use any wavelength for the ease of implementation.
In the actual implementation, we can adjust the range by adjusting the wavelengths
(or frequencies), and the detectable range (or the largest wavelength) should be chosen
3

4. Bat Algorithm
Objective function f(x), x = (x1, ..., xd)T
Initialize the bat population xi (i = 1, 2, ..., n) and vi
Define pulse frequency fi at xi
Initialize pulse rates ri and the loudness Ai
while (t <Max number of iterations)
Generate new solutions by adjusting frequency,
and updating velocities and locations/solutions [equations (2) to (4)]
if (rand > ri)
Select a solution among the best solutions
Generate a local solution around the selected best solution
end if
Generate a new solution by flying randomly
if (rand < Ai & f(xi) < f(x∗))
Accept the new solutions
Increase ri and reduce Ai
end if
Rank the bats and find the current best x∗
end while
Postprocess results and visualization
Figure 1: Pseudo code of the bat algorithm (BA).
such that it is comparable to the size of the domain of interest, and then toning down
to smaller ranges. Furthermore, we do not necessarily have to use the wavelengths
themselves, instead, we can also vary the frequency while fixing the wavelength λ. This
is because λ and f are related due to the fact λf is constant. We will use this later
approach in our implementation.
For simplicity, we can assume f ∈ [0, fmax]. We know that higher frequencies have
short wavelengths and travel a shorter distance. For bats, the typical ranges are a few
metres. The rate of pulse can simply be in the range of [0, 1] where 0 means no pulses
at all, and 1 means the maximum rate of pulse emission.
Based on these approximations and idealization, the basic steps of the Bat Algorithm
(BA) can be summarized as the pseudo code shown in Fig. 1.
3.1 Movement of Virtual Bats
In simulations, we use virtual bats naturally. We have to define the rules how their
positions xi and velocities vi in a d-dimensional search space are updated. The new
solutions xt
i and velocities vt
i at time step t are given by
fi = fmin + (fmax − fmin)β, (2)
vt
i = vt−1
i + (xt
i − x∗)fi, (3)
xt
i = xt−1
i + vt
i, (4)
where β ∈ [0, 1] is a random vector drawn from a uniform distribution. Here x∗ is the
current global best location (solution) which is located after comparing all the solutions
4

5. among all the n bats. As the product λifi is the velocity increment, we can use either fi
(or λi ) to adjust the velocity change while fixing the other factor λi (or fi), depending
on the type of the problem of interest. In our implementation, we will use fmin = 0 and
fmax = 100, depending the domain size of the problem of interest. Initially, each bat is
randomly assigned a frequency which is drawn uniformly from [fmin, fmax].
For the local search part, once a solution is selected among the current best solutions,
a new solution for each bat is generated locally using random walk
xnew = xold + ǫAt
, (5)
where ǫ ∈ [−1, 1] is a random number, while At
=<At
i > is the average loudness of all
the bats at this time step.
The update of the velocities and positions of bats have some similarity to the pro-
cedure in the standard particle swarm optimization [7] as fi essentially controls the
pace and range of the movement of the swarming particles. To a degree, BA can be
considered as a balanced combination of the standard particle swarm optimization and
the intensive local search controlled by the loudness and pulse rate.
3.2 Loudness and Pulse Emission
Furthermore, the loudness Ai and the rate ri of pulse emission have to be updated
accordingly as the iterations proceed. As the loudness usually decreases once a bat has
found its prey, while the rate of pulse emission increases, the loudness can be chosen
as any value of convenience. For example, we can use A0 = 100 and Amin = 1. For
simplicity, we can also use A0 = 1 and Amin = 0, assuming Amin = 0 means that a bat
has just found the prey and temporarily stop emitting any sound. Now we have
At+1
i = αAt
i, rt+1
i = r0
i [1 − exp(−γt)], (6)
where α and γ are constants. In fact, α is similar to the cooling factor of a cooling
schedule in the simulated annealing [9]. For any 0 < α < 1 and γ > 0, we have
At
i → 0, rt
i → r0
i , as t → ∞. (7)
In the simplicity case, we can use α = γ, and we have used α = γ = 0.9 in our
simulations. The choice of parameters requires some experimenting. Initially, each
bat should have different values of loudness and pulse emission rate, and this can be
achieved by randomization. For example, the initial loudness A0
i can typically be [1, 2],
while the initial emission rate r0
i can be around zero, or any value r0
i ∈ [0, 1] if using
(6). Their loudness and emission rates will be updated only if the new solutions are
improved, which means that these bats are moving towards the optimal solution.
4 Validation and Comparison
From the pseudo code, it is relatively straightforward to implement the Bat Algorithm
in any programming language. For the ease of visualization, we have implemented it
using Matlab for various test functions.
5

6. −5 0 5
−5
0
5
Figure 2: The paths of 25 virtual bats during 20 consecutive iterations. They converge into
(1, 1).
4.1 Benchmark Functions
There are many standard test functions for validating new algorithms. In the current
benchmark validation, we have chosen the well-known Rosenbrock’s function
f(x) =
d−1
i=1
(1 − x2
i )2
+ 100(xi+1 − x2
i )2
, −2.048 ≤ xi ≤ 2.048, (8)
and the eggcrate function
g(x, y) = x2
+ y2
+ 25(sin2
x + sin2
y), (x, y) ∈ [−2π, 2π] × [−2π, 2π]. (9)
We know that f(x) has a global minimum fmin = 0 at (1, 1) in 2D, while g(x, y) has a
global minimum gmin = 0 at (0, 0). De Jong’s standard sphere function
h(x) =
d
i=1
x2
i , −10 ≤ xi ≤ 10, (10)
has also been used. Its minimum is hmin = 0 at (0, 0, ..., 0) for any d ≥ 3.
In addition, we have also used other standard test functions for numerical global
optimization [10] such as Ackley’s function
s(x) = 20 + e − 20 exp − 0.2
1
d
d
i=1
x2
i − exp[
1
d
d
i=1
cos(2πxi)], (11)
6

7. −5
0
5
−5
0
5
0
20
40
60
80
100
−5 0 5
−5
−4
−3
−2
−1
0
1
2
3
4
5
Figure 3: The eggcrate function (left) and the locations of 40 bats in the last ten iterations
(right).
where −30 ≤ xi ≤ 30. It has the global minimum smin = 0 at (0, 0, ..., 0).
Michalewicz’s test function
f(x) = −
d
i=1
sin(xi) sin(
ix2
i
π
)
2m
, (m = 10), (12)
has d! local optima in the the domain 0 ≤ xi ≤ π where i = 1, 2, ..., d. The global
minimum is f∗ ≈ −1.801 for d = 2, while f∗ ≈ −4.6877 for d = 5.
In our implementation, we use n = 25 to 50 virtual bats, and α = 0.9. For Rosen-
brock’s 2-D banana function, the paths of 25 virtual bats during the consecutive 20 time
steps are shown in Fig. 2 where we can see that the bats converge at the global opti-
mum (1, 1). For the multimodal eggcrate function, a snapshot of the last 10 iterations
is shown in Fig. 3. Again, all bats move towards the global best (0, 0).
4.2 Comparison with Other Algorithms
In order to compare the performance of the new algorithm, we have tested it against
other heuristic algorithms, including genetic algorithms (GA) [5, 6, 11], and particle
swarm optimization (PSO) [7, 8]. There are many variants of PSO, and some variants
such as the mean PSO could perform better than the standard PSO [3]; however, the
standard PSO is by far the most popularly used. Therefore, we will also use the standard
PSO in our comparison.
There are many ways to carry out the comparison of algorithm performance, and
two obvious approaches are: to compare the numbers of function evaluations for a
given tolerance or accuracy, or to compare their accuracies for a fixed number of func-
tion evaluations. Here we will use the first approach. In our simulations, we use a
fixed tolerance ǫ ≤ 10−5
, and we run each algorithm for 100 times so that we can do
meaningful statistical analysis.
7

8. For genetic algorithms, we have used the standard version with no elitism with
the mutation probability of pm = 0.05 and crossover probability of 0.95. For particle
swarm optimization, we have also used the standard version with learning parameters
α = =
¯
2 and the inertia function I = 1 [7, 8]. The simulations have been carried out
using Matlab on a standard 3GHz desktop computer. Each run with about 10,000
function evaluations typically takes less than 5 seconds. Furthermore, we have tried to
use different population sizes from n = 10 to 250, and we found that for most problems,
n = 15 to 50 is sufficient. Therefore, we use a fixed population n = 40 for all simulations.
Table 1 shows the number of function evaluations in the form of mean ± the standard
deviation (success rate of the algorithm in finding the global optima).
Table 1: Comparison of BA with GA, and PSO.
Functions/Algorithms GA PSO BA
Multiple peaks 52124 ± 3277(98%) 3719 ± 205(97%) 1152 ± 245(100%)
Michalewicz’s (d=16) 89325 ± 7914(95%) 6922 ± 537(98%) 4752 ± 753(100%)
Rosenbrock’s (d=16) 55723 ± 8901(90%) 32756 ± 5325(98%) 7923 ± 3293(100%)
De Jong’s (d=256) 25412 ± 1237(100%) 17040 ± 1123(100%) 5273 ± 490(100%)
Schwefel’s (d=128) 227329 ± 7572(95%) 14522 ± 1275(97%) 8929 ± 729(99%)
Ackley’s (d=128) 32720 ± 3327(90%) 23407 ± 4325(92%) 6933 ± 2317(100%)
Rastrigin’s 110523 ± 5199(77%) 79491 ± 3715(90%) 12573 ± 3372(100%)
Easom’s 19239 ± 3307(92%) 17273 ± 2929(90%) 7532 ± 1702(99%)
Griewangk’s 70925 ± 7652(90%) 55970 ± 4223(92%) 9792 ± 4732(100%)
Shubert’s (18 minima) 54077 ± 4997(89%) 23992 ± 3755(92%) 11925 ± 4049(100%)
From Table 1, we can see that PSO performs much better than genetic algorithms,
while the Bat Algorithm is much superior to other algorithms in terms of accuracy and
efficiency. This is no surprising as the aim of developing the new algorithm was to try
to use the advantages of existing algorithms and other interesting feature inspired by
the fantastic behaviour of echolocation of microbats.
If we replace the variations of the frequency fi by a random parameter and setting
Ai = 0 and ri = 1, the bat algorithm essentially becomes the standard Particle Swarm
Optimization (PSO). Similarly, if we do not use the velocities, but we use fixed loudness
and rate: Ai and ri. For example, Ai = ri = 0.7, this algorithm is virtually reduced to
a simple Harmony Search (HS) as the frequency/wavelength change is essentially the
pitch adjustment, while the rate of pulse emission is similar to the harmonic acceptance
rate (here with a twist) in the harmony search algorithm [4, 15]. The current studies
implies that the proposed new algorithm is potentially more powerful and thus should
be investigated further in many applications of engineering and industrial optimization
problems.
5 Discussions
In this paper, we have successfully formulated a new Bat Algorithm for continuous
constrained optimization problems. From the formulation of the Bat Algorithm and its
implementation and comparison, we can see that it is a very promising algorithm. It is
8

9. potentially more powerful than particle swarm optimization and genetic algorithms as
well as Harmony Search. The primary reason is that BA uses a good combination of
major advantages of these algorithms in some way. Moreover, PSO and harmony search
are the special cases of the Bat Algorithm under appropriate simplifications.
In addition, the fine adjustment of the parameters α and γ can affect the convergence
rate of the bat algorithm. In fact, parameter α acts in a similar role as the cooling
schedule in the simulated annealing. Though the implementation is more complicated
than many other metaheuristic algorithms; however, it does show that it utilizes a
balanced combination of the advantages of existing successful algorithms with innovative
feature based on the echolocation behaviour of bats. New solutions are generated by
adjusting frequencies, loudness and pulse emission rates, while the proposed solution is
accepted or not depends on the quality of the solutions controlled or characterized by
loudness and pulse rate which are in turn related to the closeness or the fitness of the
locations/solution to the global optimal solution.
The exciting results suggest that more studies will highly be needed to carry out
the sensitivity analysis, to analyze the rate of algorithm convergence, and to improve
the convergence rate even further. More extensive comparison studies with a more
wide range of existing algorithms using much tough test functions in higher dimensions
will pose more challenges to the algorithms, and thus such comparisons will potentially
reveal the virtues and weakness of all the algorithms of interest.
An interesting extension will be to use different schemes of wavelength or frequency
variations instead of the current linear implementation. In addition, the rates of pulse
emission and loudness can also be varied in a more sophisticated manner. Another
extension for discrete problems is to use the time delay between pulse emission and
the echo bounced back. For example, in the travelling salesman problem, the distance
between two adjacent nodes/cities can easily be coded as time delay. As microbats use
time difference between their two ears to obtain three-dimensional information, they can
identify the type of prey and the velocity of a flying insect. Therefore, a further natural
extension to the current bat algorithm would be to use the directional echolocation and
Doppler effect, which may lead to even more interesting variants and new algorithms.
References
[1] Altringham, J. D.: Bats: Biology and Behaviour, Oxford Univesity Press, (1996).
[2] Colin, T.: The Varienty of Life. Oxford University Press, (2000).
[3] Deep, K., Bansal, J. C.: Mean particle swarm optimisation for function optimisa-
tion. Int. J. Comput. Intel. Studies, 1, 72-92 (2009).
[4] Geem, Z.W., Kim, J. H., Loganathan, G. V.: A new heuristic optimization algo-
rithm: Harmony search. Simulation, 76, 60-68 (2001).
[5] Goldberg, D.: Genetic Algorithms in Search, Optimiztion and Machine Learning.
Addison-Wesley, Reading, MA, (1989).
[6] Holland, J. H.: Adapation in Natural and Artificial Systems. University of Michigan
Press, Ann Arbor, (1975).
[7] Kennedy, J. and Eberhart, R.: Particle swarm optimization, Proc. IEEE Int. Conf.
Neural Networks. Perth, Australia, 1942-1945 (1995).
9

10. [8] Kennedy, J. and Eberhart, R., Swarm Intelligence. Academic Press, (2001).
[9] Kirkpatrick, S., Gelatt, C. D., and Vecchi, M. P.: Optimization by simulated
annealing. Science, 220, 671-680 (1983).
[10] Liang, J. J., Suganthan, P. N., Deb, K.: Novel composition test functions for
numerical global optimization. Proc. IEEE Int. Swarm Intel. Symp., 68-75 (2005).
[11] Mitchell, M.: An Introduction to Genetic Algorithms. MIT Press, (1998).
[12] Richardson, P.: Bats. Natural History Museum, London, (2008).
[13] Richardson, P.: The secrete life of bats. http://www.nhm.ac.uk
[14] Yang, X.-S.: Nature-inspired Metaheuristic Algorithms. Luniver Press, (2008).
[15] Yang, X.-S.: Harmony search as a metaheuristic algorithm, in: Music-Inspired
Harmony Search Algorithm: Theory and Applications (eds Z. W. Geem). Springer,
pp.1-14, (2009).
10